[BenjaminTay]

Fixes #295.

The boundary discrepancy reported in the issue (~1 ns) and its float64 frame-arithmetic example (~1.11e-16 s) can make chord.directional_hamming_distance raise ValueError, preventing chord evaluation from returning scores.

This change treats adjacent intervals as touching only when their overlap is at most 10 ns and 1 ppm of the shorter neighboring duration. It moves the later start to the preceding end in a local copy. Reference and estimated intervals use the same adjustment before the adjacent-overlap check and scoring; caller arrays and all positive gaps are preserved.

Engineering tradeoffs

• 10 ns gives a tenfold margin over the boundary discrepancy reported in the issue. 1 ppm limits duration changes and protects short intervals from collapse. These are engineering choices proposed for maintainer review, not established tolerances or machine-epsilon bounds. Unlike whole-array millisecond rounding, adjustments remain local.
• Genuine micro-overlaps within both limits are also absorbed: the values alone cannot distinguish them from numerical artifacts.
• The limits do not adapt to dtype or timestamp scale. The original float64 examples are covered; the analogous float32 frame example can overlap by ~5.96e-8 seconds and still raises.

Validation

The 11 added cases cover reported inputs, both scoring directions, read-only inputs, preserved gaps, and rejected material overlaps/short intervals. Against the unmodified base they give 4 failed / 7 passed; with this patch, 11 passed.

Full local runs used the same macOS/Python 3.12 environment and base fe73b3533737814f83dbd9739f06e90f5f82f758, running pytest --mpl-results-path=mpl_image_results/ from tests/ with default options:

Source	Passed	Failed	Skipped	Xfailed	Xpassed
Unmodified base	516	20	3	175	1
This patch	527	20	3	175	1

The full suite did not pass. Both runs exited 1 with the same 20 test_display.py image failures: matching test IDs and RMS values, and 60 byte-identical comparison PNGs. This comparison found no additional failures; it does not replace upstream CI. Image baselines were not changed.

[craffel]

@bmcfee what do you think of this approach for fixing #295 which you had originally raised?

[bmcfee]

It seems pretty reasonable to me. I'll also loop in @tomxi - would this intersect with your exact segment code in a useful way?